func TestRaftQuorumRecovery(t *testing.T) {
t.Parallel()
// Bring up a 5 nodes cluster
nodes, clockSource := raftutils.NewRaftCluster(t, tc)
raftutils.AddRaftNode(t, clockSource, nodes, tc)
raftutils.AddRaftNode(t, clockSource, nodes, tc)
defer raftutils.TeardownCluster(t, nodes)
// Lose a majority
for i := uint64(1); i <= 3; i++ {
nodes[i].Server.Stop()
nodes[i].Shutdown()
}
raftutils.AdvanceTicks(clockSource, 5)
// Restore the majority by restarting node 3
nodes[3] = raftutils.RestartNode(t, clockSource, nodes[3], false)
delete(nodes, 1)
delete(nodes, 2)
raftutils.WaitForCluster(t, clockSource, nodes)
// Propose a value
value, err := raftutils.ProposeValue(t, raftutils.Leader(nodes))
assert.NoError(t, err)
for _, node := range nodes {
raftutils.CheckValue(t, clockSource, node, value)
}
}
func TestRaftFollowerLeave(t *testing.T) {
t.Parallel()
// Bring up a 5 nodes cluster
nodes, clockSource := raftutils.NewRaftCluster(t, tc)
raftutils.AddRaftNode(t, clockSource, nodes, tc)
raftutils.AddRaftNode(t, clockSource, nodes, tc)
defer raftutils.TeardownCluster(t, nodes)
// Node 5 leaves the cluster
// Use gRPC instead of calling handler directly because of
// authorization check.
cc, err := dial(nodes[1], nodes[1].Address)
assert.NoError(t, err)
raftClient := api.NewRaftMembershipClient(cc)
defer cc.Close()
ctx, _ := context.WithTimeout(context.Background(), 10*time.Second)
resp, err := raftClient.Leave(ctx, &api.LeaveRequest{Node: &api.RaftMember{RaftID: nodes[5].Config.ID}})
assert.NoError(t, err, "error sending message to leave the raft")
assert.NotNil(t, resp, "leave response message is nil")
raftutils.ShutdownNode(nodes[5])
delete(nodes, 5)
raftutils.WaitForPeerNumber(t, clockSource, nodes, 4)
// Propose a value
value, err := raftutils.ProposeValue(t, nodes[1], DefaultProposalTime)
assert.NoError(t, err, "failed to propose value")
// Value should be replicated on every node
raftutils.CheckValue(t, clockSource, nodes[1], value)
assert.Len(t, nodes[1].GetMemberlist(), 4)
raftutils.CheckValue(t, clockSource, nodes[2], value)
assert.Len(t, nodes[2].GetMemberlist(), 4)
raftutils.CheckValue(t, clockSource, nodes[3], value)
assert.Len(t, nodes[3].GetMemberlist(), 4)
raftutils.CheckValue(t, clockSource, nodes[4], value)
assert.Len(t, nodes[4].GetMemberlist(), 4)
}
func TestRaftQuorumFailure(t *testing.T) {
t.Parallel()
// Bring up a 5 nodes cluster
nodes, clockSource := raftutils.NewRaftCluster(t, tc)
raftutils.AddRaftNode(t, clockSource, nodes, tc)
raftutils.AddRaftNode(t, clockSource, nodes, tc)
defer raftutils.TeardownCluster(t, nodes)
// Lose a majority
for i := uint64(3); i <= 5; i++ {
nodes[i].Server.Stop()
nodes[i].Stop()
}
// Propose a value
_, err := raftutils.ProposeValue(t, nodes[1])
assert.Error(t, err)
// The value should not be replicated, we have no majority
raftutils.CheckNoValue(t, clockSource, nodes[2])
raftutils.CheckNoValue(t, clockSource, nodes[1])
}
func TestRaftNewNodeGetsData(t *testing.T) {
t.Parallel()
// Bring up a 3 node cluster
nodes, clockSource := raftutils.NewRaftCluster(t, tc)
defer raftutils.TeardownCluster(t, nodes)
// Propose a value
value, err := raftutils.ProposeValue(t, nodes[1])
assert.NoError(t, err, "failed to propose value")
// Add a new node
raftutils.AddRaftNode(t, clockSource, nodes, tc)
time.Sleep(500 * time.Millisecond)
// Value should be replicated on every node
for _, node := range nodes {
raftutils.CheckValue(t, clockSource, node, value)
assert.Equal(t, len(node.GetMemberlist()), 4)
}
}
func TestStress(t *testing.T) {
t.Parallel()
// Bring up a 5 nodes cluster
nodes, clockSource := raftutils.NewRaftCluster(t, tc)
raftutils.AddRaftNode(t, clockSource, nodes, tc)
raftutils.AddRaftNode(t, clockSource, nodes, tc)
defer raftutils.TeardownCluster(t, nodes)
// number of nodes that are running
nup := len(nodes)
// record of nodes that are down
idleNodes := map[int]struct{}{}
// record of ids that proposed successfully or time-out
pIDs := []string{}
leader := -1
for iters := 0; iters < 1000; iters++ {
// keep proposing new values and killing leader
for i := 1; i <= 5; i++ {
if nodes[uint64(i)] != nil {
id := strconv.Itoa(iters)
_, err := raftutils.ProposeValue(t, nodes[uint64(i)], id)
if err == nil {
pIDs = append(pIDs, id)
// if propose successfully, at least there are 3 running nodes
assert.True(t, nup >= 3)
// only leader can propose value
assert.True(t, leader == i || leader == -1)
// update leader
leader = i
break
} else if strings.Contains(err.Error(), "context deadline exceeded") {
// though it's timing out, we still record this value
// for it may be proposed successfully and stored in Raft some time later
pIDs = append(pIDs, id)
}
}
}
if rand.Intn(100) < 10 {
// increase clock to make potential election finish quickly
clockSource.Increment(200 * time.Millisecond)
time.Sleep(10 * time.Millisecond)
} else {
ms := rand.Intn(10)
clockSource.Increment(time.Duration(ms) * time.Millisecond)
}
if leader != -1 {
// if propose successfully, try to kill a node in random
s := rand.Intn(5) + 1
if _, ok := idleNodes[s]; !ok {
id := uint64(s)
nodes[id].Server.Stop()
nodes[id].Shutdown()
idleNodes[s] = struct{}{}
nup -= 1
if s == leader {
// leader is killed
leader = -1
}
}
}
if nup < 3 {
// if quorum is lost, try to bring back a node
s := rand.Intn(5) + 1
if _, ok := idleNodes[s]; ok {
id := uint64(s)
nodes[id] = raftutils.RestartNode(t, clockSource, nodes[id], false)
delete(idleNodes, s)
nup++
}
}
}
// bring back all nodes and propose the final value
for i := range idleNodes {
id := uint64(i)
nodes[id] = raftutils.RestartNode(t, clockSource, nodes[id], false)
}
raftutils.WaitForCluster(t, clockSource, nodes)
id := strconv.Itoa(1000)
val, err := raftutils.ProposeValue(t, raftutils.Leader(nodes), id)
assert.NoError(t, err, "failed to propose value")
pIDs = append(pIDs, id)
// increase clock to make cluster stable
time.Sleep(500 * time.Millisecond)
clockSource.Increment(500 * time.Millisecond)
ids, values := raftutils.GetAllValuesOnNode(t, clockSource, nodes[1])
// since cluster is stable, final value must be in the raft store
find := false
for _, value := range values {
if reflect.DeepEqual(value, val) {
find = true
break
}
}
assert.True(t, find)
// all nodes must have the same value
raftutils.CheckValuesOnNodes(t, clockSource, nodes, ids, values)
// ids should be a subset of pIDs
for _, id := range ids {
find = false
for _, pid := range pIDs {
if id == pid {
find = true
break
}
}
assert.True(t, find)
}
}
func TestRaftSnapshot(t *testing.T) {
t.Parallel()
// Bring up a 3 node cluster
nodes, clockSource := raftutils.NewRaftCluster(t, tc, &api.RaftConfig{SnapshotInterval: 9, LogEntriesForSlowFollowers: 0})
defer raftutils.TeardownCluster(t, nodes)
nodeIDs := []string{"id1", "id2", "id3", "id4", "id5", "id6", "id7", "id8", "id9", "id10", "id11", "id12"}
values := make([]*api.Node, len(nodeIDs))
snapshotFilenames := make(map[uint64]string, 4)
// Propose 3 values
var err error
for i, nodeID := range nodeIDs[:3] {
values[i], err = raftutils.ProposeValue(t, nodes[1], DefaultProposalTime, nodeID)
assert.NoError(t, err, "failed to propose value")
}
// None of the nodes should have snapshot files yet
for _, node := range nodes {
dirents, err := ioutil.ReadDir(filepath.Join(node.StateDir, "snap"))
assert.NoError(t, err)
assert.Len(t, dirents, 0)
}
// Check all nodes have all the data.
// This also acts as a synchronization point so that the next value we
// propose will arrive as a separate message to the raft state machine,
// and it is guaranteed to have the right cluster settings when
// deciding whether to create a new snapshot.
raftutils.CheckValuesOnNodes(t, clockSource, nodes, nodeIDs[:3], values)
// Propose a 4th value
values[3], err = raftutils.ProposeValue(t, nodes[1], DefaultProposalTime, nodeIDs[3])
assert.NoError(t, err, "failed to propose value")
// All nodes should now have a snapshot file
for nodeID, node := range nodes {
assert.NoError(t, raftutils.PollFunc(clockSource, func() error {
dirents, err := ioutil.ReadDir(filepath.Join(node.StateDir, "snap"))
if err != nil {
return err
}
if len(dirents) != 1 {
return fmt.Errorf("expected 1 snapshot, found %d", len(dirents))
}
snapshotFilenames[nodeID] = dirents[0].Name()
return nil
}))
}
// Add a node to the cluster
raftutils.AddRaftNode(t, clockSource, nodes, tc)
// It should get a copy of the snapshot
assert.NoError(t, raftutils.PollFunc(clockSource, func() error {
dirents, err := ioutil.ReadDir(filepath.Join(nodes[4].StateDir, "snap"))
if err != nil {
return err
}
if len(dirents) != 1 {
return fmt.Errorf("expected 1 snapshot, found %d on new node", len(dirents))
}
snapshotFilenames[4] = dirents[0].Name()
return nil
}))
// It should know about the other nodes
stripMembers := func(memberList map[uint64]*api.RaftMember) map[uint64]*api.RaftMember {
raftNodes := make(map[uint64]*api.RaftMember)
for k, v := range memberList {
raftNodes[k] = &api.RaftMember{
RaftID: v.RaftID,
Addr: v.Addr,
}
}
return raftNodes
}
assert.Equal(t, stripMembers(nodes[1].GetMemberlist()), stripMembers(nodes[4].GetMemberlist()))
// All nodes should have all the data
raftutils.CheckValuesOnNodes(t, clockSource, nodes, nodeIDs[:4], values)
// Propose more values to provoke a second snapshot
for i := 4; i != len(nodeIDs); i++ {
values[i], err = raftutils.ProposeValue(t, nodes[1], DefaultProposalTime, nodeIDs[i])
assert.NoError(t, err, "failed to propose value")
}
// All nodes should have a snapshot under a *different* name
for nodeID, node := range nodes {
assert.NoError(t, raftutils.PollFunc(clockSource, func() error {
dirents, err := ioutil.ReadDir(filepath.Join(node.StateDir, "snap"))
if err != nil {
return err
}
if len(dirents) != 1 {
return fmt.Errorf("expected 1 snapshot, found %d on node %d", len(dirents), nodeID)
}
if dirents[0].Name() == snapshotFilenames[nodeID] {
return fmt.Errorf("snapshot %s did not get replaced", snapshotFilenames[nodeID])
}
return nil
}))
}
// All nodes should have all the data
raftutils.CheckValuesOnNodes(t, clockSource, nodes, nodeIDs, values)
}
func TestCanRemoveMember(t *testing.T) {
nodes, clockSource := raftutils.NewRaftCluster(t, tc)
defer raftutils.TeardownCluster(t, nodes)
// Stop node 2 and node 3 (2 nodes out of 3)
nodes[2].Server.Stop()
nodes[2].Shutdown()
nodes[3].Server.Stop()
nodes[3].Shutdown()
// Node 2 and Node 3 should be listed as Unreachable
assert.NoError(t, raftutils.PollFunc(clockSource, func() error {
members := nodes[1].GetMemberlist()
if len(members) != 3 {
return fmt.Errorf("expected 3 nodes, got %d", len(members))
}
if members[nodes[2].Config.ID].Status.Reachability == api.RaftMemberStatus_REACHABLE {
return fmt.Errorf("expected node 2 to be unreachable")
}
if members[nodes[3].Config.ID].Status.Reachability == api.RaftMemberStatus_REACHABLE {
return fmt.Errorf("expected node 3 to be unreachable")
}
return nil
}))
// Removing all nodes should fail
for i := 1; i <= 3; i++ {
ctx, _ := context.WithTimeout(context.Background(), 10*time.Second)
err := nodes[1].RemoveMember(ctx, uint64(i))
assert.Error(t, err)
assert.Equal(t, err, raft.ErrCannotRemoveMember)
members := nodes[1].GetMemberlist()
assert.Equal(t, len(members), 3)
}
// Restart node 2 and node 3
nodes[2] = raftutils.RestartNode(t, clockSource, nodes[2], false)
nodes[3] = raftutils.RestartNode(t, clockSource, nodes[3], false)
raftutils.WaitForCluster(t, clockSource, nodes)
// Node 2 and Node 3 should be listed as Reachable
assert.NoError(t, raftutils.PollFunc(clockSource, func() error {
members := nodes[1].GetMemberlist()
if len(members) != 3 {
return fmt.Errorf("expected 3 nodes, got %d", len(members))
}
if members[nodes[2].Config.ID].Status.Reachability == api.RaftMemberStatus_UNREACHABLE {
return fmt.Errorf("expected node 2 to be reachable")
}
if members[nodes[3].Config.ID].Status.Reachability == api.RaftMemberStatus_UNREACHABLE {
return fmt.Errorf("expected node 3 to be reachable")
}
return nil
}))
// Stop Node 3 (1 node out of 3)
nodes[3].Server.Stop()
nodes[3].Shutdown()
// Node 3 should be listed as Unreachable
assert.NoError(t, raftutils.PollFunc(clockSource, func() error {
members := nodes[1].GetMemberlist()
if len(members) != 3 {
return fmt.Errorf("expected 3 nodes, got %d", len(members))
}
if members[nodes[3].Config.ID].Status.Reachability == api.RaftMemberStatus_REACHABLE {
return fmt.Errorf("expected node 3 to be unreachable")
}
return nil
}))
// Removing node 2 should fail (this would break the quorum)
ctx, _ := context.WithTimeout(context.Background(), 10*time.Second)
err := nodes[1].RemoveMember(ctx, nodes[2].Config.ID)
assert.Error(t, err)
assert.Equal(t, err, raft.ErrCannotRemoveMember)
members := nodes[1].GetMemberlist()
assert.Equal(t, len(members), 3)
// Removing node 3 works fine because it is already unreachable
ctx, _ = context.WithTimeout(context.Background(), 10*time.Second)
err = nodes[1].RemoveMember(ctx, nodes[3].Config.ID)
assert.NoError(t, err)
members = nodes[1].GetMemberlist()
assert.Nil(t, members[nodes[3].Config.ID])
assert.Equal(t, len(members), 2)
// Add back node 3
raftutils.ShutdownNode(nodes[3])
delete(nodes, 3)
raftutils.AddRaftNode(t, clockSource, nodes, tc)
// Node 2 and Node 3 should be listed as Reachable
assert.NoError(t, raftutils.PollFunc(clockSource, func() error {
members := nodes[1].GetMemberlist()
if len(members) != 3 {
return fmt.Errorf("expected 3 nodes, got %d", len(members))
}
if members[nodes[2].Config.ID].Status.Reachability != api.RaftMemberStatus_REACHABLE {
return fmt.Errorf("expected node 2 to be reachable")
}
if members[nodes[3].Config.ID].Status.Reachability != api.RaftMemberStatus_REACHABLE {
return fmt.Errorf("expected node 3 to be reachable")
}
return nil
}))
// Removing node 3 should succeed
ctx, _ = context.WithTimeout(context.Background(), 10*time.Second)
err = nodes[1].RemoveMember(ctx, nodes[3].Config.ID)
assert.NoError(t, err)
members = nodes[1].GetMemberlist()
assert.Nil(t, members[nodes[3].Config.ID])
assert.Equal(t, len(members), 2)
// Removing node 2 should succeed
ctx, _ = context.WithTimeout(context.Background(), 10*time.Second)
err = nodes[1].RemoveMember(ctx, nodes[2].Config.ID)
assert.NoError(t, err)
members = nodes[1].GetMemberlist()
assert.Nil(t, members[nodes[2].Config.ID])
assert.Equal(t, len(members), 1)
}
func TestListManagerNodes(t *testing.T) {
t.Parallel()
tc := cautils.NewTestCA(nil)
defer tc.Stop()
ts := newTestServer(t)
defer ts.Stop()
nodes, clockSource := raftutils.NewRaftCluster(t, tc)
defer raftutils.TeardownCluster(t, nodes)
// Create a node object for each of the managers
assert.NoError(t, nodes[1].MemoryStore().Update(func(tx store.Tx) error {
assert.NoError(t, store.CreateNode(tx, &api.Node{ID: nodes[1].SecurityConfig.ClientTLSCreds.NodeID()}))
assert.NoError(t, store.CreateNode(tx, &api.Node{ID: nodes[2].SecurityConfig.ClientTLSCreds.NodeID()}))
assert.NoError(t, store.CreateNode(tx, &api.Node{ID: nodes[3].SecurityConfig.ClientTLSCreds.NodeID()}))
return nil
}))
// Assign one of the raft node to the test server
ts.Server.raft = nodes[1].Node
ts.Server.store = nodes[1].MemoryStore()
// There should be 3 reachable managers listed
r, err := ts.Client.ListNodes(context.Background(), &api.ListNodesRequest{})
assert.NoError(t, err)
assert.NotNil(t, r)
managers := getMap(t, r.Nodes)
assert.Len(t, ts.Server.raft.GetMemberlist(), 3)
assert.Len(t, r.Nodes, 3)
// Node 1 should be the leader
for i := 1; i <= 3; i++ {
if i == 1 {
assert.True(t, managers[nodes[uint64(i)].Config.ID].Leader)
continue
}
assert.False(t, managers[nodes[uint64(i)].Config.ID].Leader)
}
// All nodes should be reachable
for i := 1; i <= 3; i++ {
assert.Equal(t, api.RaftMemberStatus_REACHABLE, managers[nodes[uint64(i)].Config.ID].Reachability)
}
// Add two more nodes to the cluster
raftutils.AddRaftNode(t, clockSource, nodes, tc)
raftutils.AddRaftNode(t, clockSource, nodes, tc)
raftutils.WaitForCluster(t, clockSource, nodes)
// Add node entries for these
assert.NoError(t, nodes[1].MemoryStore().Update(func(tx store.Tx) error {
assert.NoError(t, store.CreateNode(tx, &api.Node{ID: nodes[4].SecurityConfig.ClientTLSCreds.NodeID()}))
assert.NoError(t, store.CreateNode(tx, &api.Node{ID: nodes[5].SecurityConfig.ClientTLSCreds.NodeID()}))
return nil
}))
// There should be 5 reachable managers listed
r, err = ts.Client.ListNodes(context.Background(), &api.ListNodesRequest{})
assert.NoError(t, err)
assert.NotNil(t, r)
managers = getMap(t, r.Nodes)
assert.Len(t, ts.Server.raft.GetMemberlist(), 5)
assert.Len(t, r.Nodes, 5)
for i := 1; i <= 5; i++ {
assert.Equal(t, api.RaftMemberStatus_REACHABLE, managers[nodes[uint64(i)].Config.ID].Reachability)
}
// Stops 2 nodes
nodes[4].Server.Stop()
nodes[4].ShutdownRaft()
nodes[5].Server.Stop()
nodes[5].ShutdownRaft()
// Node 4 and Node 5 should be listed as Unreachable
assert.NoError(t, raftutils.PollFunc(clockSource, func() error {
r, err = ts.Client.ListNodes(context.Background(), &api.ListNodesRequest{})
if err != nil {
return err
}
managers = getMap(t, r.Nodes)
if len(r.Nodes) != 5 {
return fmt.Errorf("expected 5 nodes, got %d", len(r.Nodes))
}
if managers[nodes[4].Config.ID].Reachability == api.RaftMemberStatus_REACHABLE {
return fmt.Errorf("expected node 4 to be unreachable")
}
if managers[nodes[5].Config.ID].Reachability == api.RaftMemberStatus_REACHABLE {
return fmt.Errorf("expected node 5 to be unreachable")
}
return nil
}))
// Restart the 2 nodes
nodes[4] = raftutils.RestartNode(t, clockSource, nodes[4], false)
nodes[5] = raftutils.RestartNode(t, clockSource, nodes[5], false)
raftutils.WaitForCluster(t, clockSource, nodes)
assert.Len(t, ts.Server.raft.GetMemberlist(), 5)
// All the nodes should be reachable again
assert.NoError(t, raftutils.PollFunc(clockSource, func() error {
r, err = ts.Client.ListNodes(context.Background(), &api.ListNodesRequest{})
if err != nil {
return err
}
managers = getMap(t, r.Nodes)
for i := 1; i <= 5; i++ {
if managers[nodes[uint64(i)].Config.ID].Reachability != api.RaftMemberStatus_REACHABLE {
return fmt.Errorf("node %x is unreachable", nodes[uint64(i)].Config.ID)
}
}
return nil
}))
// Switch the raft node used by the server
ts.Server.raft = nodes[2].Node
// Stop node 1 (leader)
nodes[1].Server.Stop()
nodes[1].ShutdownRaft()
newCluster := map[uint64]*raftutils.TestNode{
2: nodes[2],
3: nodes[3],
4: nodes[4],
5: nodes[5],
}
// Wait for the re-election to occur
raftutils.WaitForCluster(t, clockSource, newCluster)
// Node 1 should not be the leader anymore
assert.NoError(t, raftutils.PollFunc(clockSource, func() error {
r, err = ts.Client.ListNodes(context.Background(), &api.ListNodesRequest{})
if err != nil {
return err
}
managers = getMap(t, r.Nodes)
if managers[nodes[1].Config.ID].Leader {
return fmt.Errorf("expected node 1 not to be the leader")
}
if managers[nodes[1].Config.ID].Reachability == api.RaftMemberStatus_REACHABLE {
return fmt.Errorf("expected node 1 to be unreachable")
}
return nil
}))
// Restart node 1
nodes[1].ShutdownRaft()
nodes[1] = raftutils.RestartNode(t, clockSource, nodes[1], false)
raftutils.WaitForCluster(t, clockSource, nodes)
// Ensure that node 1 is not the leader
assert.False(t, managers[nodes[uint64(1)].Config.ID].Leader)
// Check that another node got the leader status
var leader uint64
leaderCount := 0
for i := 1; i <= 5; i++ {
if managers[nodes[uint64(i)].Config.ID].Leader {
leader = nodes[uint64(i)].Config.ID
leaderCount++
}
}
// There should be only one leader after node 1 recovery and it
// should be different than node 1
assert.Equal(t, 1, leaderCount)
assert.NotEqual(t, leader, nodes[1].Config.ID)
}